Methods and apparatus for real-time interaction analysis in call centers

• US 9,015,046 B2
• Filed: 06/10/2010
• Issued: 04/21/2015
• Est. Priority Date: 06/10/2010
• Status: Active Grant

• Current Assignee: Nice Ltd
• Original Assignee: Nice Systems Limited (Nice Ltd)
• Inventors: Pereg, Oren, Wasserblat, Moshe, Lubowich, Yuval, Laperdon, Ronen, Shapira, Dori, Feigin, Vladislav, Fox-Kahana, Oz
• Primary Examiner(s): Shah, Paras D
• Application Number:
  US12/797,618
• Publication Number:
  US 20110307257A1
• Time in Patent Office: 1,776 Days
• Field of Search: 704/231, 704/246, 704/251, 704/235, 704/257, 704/254, 704/270.1, 379/88.01, 37926501-26507
• US Class Current: 704/251
• CPC Class Codes: G06Q 10/063 Operations research, analysis or management
• CPC Class Codes: G06Q 30/01 Customer relationship services
• CPC Class Codes: G06Q 30/0202 Market predictions or forecasting for commercial activities
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 2015/081 Search algorithms, e.g. Baum-Welch or Viterbi
• CPC Class Codes: G10L 2015/085 Methods for reducing search complexity, pruning

Abstract

A method and system for indicating in real time that an interaction is associated with a problem or issue, comprising: receiving a segment of an interaction in which a representative of the organization participates; extracting a feature from the segment; extracting a global feature associated with the interaction; aggregating the feature and the global feature; and classifying the segment or the interaction in association with the problem or issue by applying a model to the feature and the global feature. The method and system may also use features extracted from earlier segments within the interaction. The method and system can also evaluate the model based on features extracted from training interactions and manual tagging assigned to the interactions or segments thereof.

Claims

1. A computerized method for indicating in real time that an interaction in which a representative of an organization participates is associated with a problem or issue, comprising:

capturing the interaction in a storage device by a computing platform comprising a processing apparatus, the interaction comprises at least a speech;

receiving a segment of the captured interaction;

extracting a feature from the segment wherein the feature is a word spoken in the segment of the interaction obtained by a speech to text mechanism comprising hardware;

extracting a global feature associated with the whole interaction in progress;

extracting features from segments of the interaction that are earlier to the segment;

aggregating the feature and the global feature and features similar to the feature that are related to the segments;

classifying the segment of the interaction in association with the problem or issue by applying a model to the feature and the global feature and the features similar to the feature that are related to the segments,

wherein extracting the feature from the segment comprises;

generating a phoneme graph from the segment of the interaction;

flushing a part of the graph by removing graph edges that have no continuity or do not lead to a final state of a HMM, wherein the edges represent candidate phonemes between two time points;

coarsely examining the flushed graph and determining a point in the flushed graph of high probability, wherein the point comprises a most probable sub-path of a path;

determining a limited time segment window based on the most probable sub-path that covers the sub-paths above or below the most probable sub-path; and

performing a thorough search only over the limited time segment window, wherein the thorough search comprises all edges which appear in the limited time window,

determining one or more words based on the thorough search,

wherein the recited operations are carried out in real time by a computing platform executing one or more computer applications.

2. The method of claim 1 further comprising determining an action to be taken when the classification indicated that the segment of the interaction is associated with the problem or issue.

3. The method of claim 1 further comprising,:

receiving a training corpus;

framing, an interaction of the test corpus into segments;

extracting features from the segments; and

evaluating, the model based on the features.

4. The method of claim 1 wherein the feature relates to the segment and is selected from the group consisting of:

a spotted word;

a feature or indication associated with emotion;

an emotion score that represents the probability of emotion to exist within the segment;

a talk analysis parameters;

an agent burst position;

a customer burst position;

a number of bursts;

agent talk percentage;

customer talk percentage;

a silence duration;

a word extracted by a speech to text engine;

number of holds;

hold duration;

hold position;

number of transfers;

part of speech tagging or stemming;

segment position within the interaction in absolute time, segment position within the interaction in number of words;

speaker side within the segment; and

average duration of silence between words in the segment.

5. The method of claim 1 wherein the global feature is selected from the group consisting of:

average agent activity within the interaction;

average customer activity within the interaction;

average silence;

number of bursts per participant side;

total burst duration per participant side;

an emotion detection feature;

total probability score for an emotion within the interaction;

number of emotion bursts;

total emotion burst duration;

a stemmed word, a stemmed keyphrase;

number of word repetitions; and

number of keyphrase repetitions.

7. The method of claim 2 wherein the action is selected from the group consisting of:

popping a message on a display device used by a person;

generating an alert;

muting the interaction;

conferencing the interaction; and

adding data to a statistical storage.

8. The method of claim 1 wherein the problem or issues is selected from the group consisting of:

customer dissatisfaction;

churn prediction, sales assistance; and

fraud detection.

6. The method of claim wherein the aggregation step aggregates a feature related to a second segment preceding the segment, the feature selected, from the group consisting of:

a spotted word;

a feature or indication associated with emotion;

an emotion score that represents the probability of emotion to exist within the segment;

a talk analysis parameters;

an agent burst position;

a customer burst position;

a number of bursts;

agent talk percentage;

customer talk percentage;

a silence duration;

a word extracted by a speech to text engine;

number of holds;

hold duration;

hold position;

number of transfers;

part of speech tagging or stemming;

segment position within the interaction in absolute time, segment position within the interaction in number of words;

speaker side within the segment; and

average duration of silence between words in the segment.

9. The method of claim wherein the segment of the interaction overlaps with a previous or following segment of the interaction.